Lippert explained that the choice of machine is application driven. The system is used by for running demanding applications. Hence, a Beowulf cluster would not be sufficient. Lippert stressed that the machine is a fully fledged parallel supercomputer. Apart from the Alpha processors, which are fast, a fast network interconnect was needed. For this Myrinet was chosen, which has a much higher bandwidth and lower latency that the Fast-Ethernet used in Beowulf clusters, with performance in the Cray T3E range. ParaStation software is used to create a one-system view for the user, of this Myrinet connected collection of processors. As operating system they did choose Linux: SuSe version 6.4.
Job management system is the Portable Batch System PBS. Because the cluster has to be managed efficiently, the Compaq Cluster Management Utility software was installed. This takes care of node management, user administration, and monitoring. At Wuppertal they additionally developed LookingGlass software to couple ALiCE to rendering software at the desktop.
Once the machine has been upgraded to 128 nodes at 625 MHz, the machine will have a peak of 150 Gflop/s. Memory size of the nodes will add up to 32 Gbyte; the disk space of ALiCE will be 1.2 Tbyte. The machine will have a power consumption of 15 KWh. Lippert expects the machine will run the LinPack benchmark at 85 Gflop/s.
ALiCE was tested on several application codes including a Quantum Field Code for which the machine performed the same as the Cray T3E.
Overall ALiCE is doing very well at Wuppertal. There are a few things to be improved, said Lippert. Currently I/O is a bottleneck. A parallel file system could be a solution here.